Enhancing accuracy of one-dimensional characteristic predictions for axial compressors using deep learning

To enhance the accuracy of one-dimensional (1D) characteristic predictions for multistage axial flow compressors and improve design efficiency, this paper presents a deviation angle prediction model based on deep learning, utilizing data generated from Computational Fluid Dynamics (CFD) simulations. This model was integrated into a compressor characteristic calculation program. A CFD simulation dataset was created with National Advisory Committee for Aeronautics 65-series (NACA65) blade profiles, parameterized through a Latin hypercube design. After solving blade cascade flow fields, deviation angles under various conditions were obtained, establishing a mapping between design variables and deviation angles using deep learning. Test results showed a correlation coefficient of 0.9978 and a mean absolute error of 0.0785°. The surrogate model was embedded into the axial flow compressor 1D calculation program (HARIKA), replacing the original deviation angle model. Performance calculations on transonic two-stage and high-subsonic eight-stage compressors at different speeds demonstrated that the updated HARIKA program provided predictions closer to experimental values at high speeds, though slightly overestimated at lower speeds. These results confirm the model’s accuracy and practicality, suggesting that the improved HARIKA algorithm has potential for engineering applications in predicting the aerodynamic performance of multistage axial flow compressors.